wav2vec2-large-xlsr-korean-demo3

This model is a fine-tuned version of NX2411/wav2vec2-large-xlsr-korean-demo-no-LM on the zeroth_korean dataset. It achieves the following results on the evaluation set:

• Loss: 0.8265
• Wer: 0.5090

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.6157	2.6	400	0.6686	0.6386
0.4643	5.19	800	0.7036	0.6086
0.3038	7.79	1200	0.6960	0.5817
0.2229	10.39	1600	0.7358	0.5571
0.178	12.99	2000	0.8221	0.5636
0.153	15.58	2400	0.8575	0.5691
0.129	18.18	2800	0.7809	0.5297
0.1141	20.78	3200	0.8077	0.5441
0.0994	23.38	3600	0.8087	0.5209
0.0917	25.97	4000	0.8176	0.5149
0.0823	28.57	4400	0.8265	0.5090

Framework versions

• Transformers 4.21.1
• Pytorch 1.12.1+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1